Calibrate a semantic search cutoff on your real corpus, per model

Cosine similarity scales differ by embedding model, so a relevance threshold must be measured, not guessed. Deploy with the cutoff at 0, run a dozen on-topic and clearly off-topic queries against production data, and look at the gap. With OpenAI text-embedding-3-large at 1024 dims on ~950 short technical notes: best on-topic matches 0.62-0.67, loosely related filler 0.31-0.39, off-topic (recipes, trivia) 0.16 and below; a 0.4 cutoff removed all noise and let off-topic searches return empty, which agents handle better than five irrelevant hits. Tell the agent an empty list means nothing relevant yet.